Admission is not eviction

An application may consider a response cacheable, an admission policy may reject it, a replacement policy may choose a victim, and an allocator may still lack room for its size class. Expiration separately determines whether an entry remains eligible for reuse by age. None of these is equivalent to the others. A miss counter cannot identify which decision caused absence, and an eviction count does not tell whether rejecting the incoming object would have been better.

Pure LRU favors recent access, while frequency-based admission can compare a candidate's estimated recent popularity with that of a proposed victim. Estimates can be approximate and aged so old popularity does not dominate forever. A policy tuned for stationary traffic can struggle after a workload shift. Avoid claiming one algorithm universally wins: the arrival distribution, object sizes, concurrency costs, and value of avoiding each origin fetch determine the useful comparison.

Suppose a cache holds a hot set that fits exactly, then a one-time scan accesses many new objects. An admit-everything LRU policy can evict the entire hot set while recording no future benefit from the scan entries. Rejecting some one-hit candidates can protect reuse, but a genuinely new popular object must eventually gain admission. Windowed or adaptive designs balance those pressures. Replay a trace with phase changes rather than optimizing only the final steady segment.

Object sizes matter as well as counts. One large object can displace several smaller valuable ones, and a high request hit rate can coexist with poor byte savings or expensive remaining misses. Metadata, allocator overhead, and policy bookkeeping consume memory and CPU. Compare actual resident bytes and origin work at the same total resource budget. A policy with a slightly higher simulated hit ratio may lose in production if its coordination cost overwhelms the saved fetches.

Separate trace replay from live outcome measurement. A replay can estimate a defined policy's decisions for a recorded stream, but the trace may change when latency or admission changes. Warmup, expiration, invalidations, and concurrent requests need explicit treatment. In a live comparison retain correctness and freshness checks, observe origin cost, and sample policy CPU overhead. Do not compare a count-bounded simulation with a byte-bounded production cache as if their capacities were identical.

The cache lab implements only fully associative LRU and admits every miss. Reset and set both C=4, s=4, predict twelve hits in sixteen initial requests, then Step. Reset and re-set C=3, s=4; predict zero hits because the four-address cycle exceeds three entries. Its address stream is (i×s) mod 16 and capped at 4096 requests. There is no admission switch, variable size, TTL, or TinyLFU estimator. Explain how rejecting candidates would require a different trace and policy state rather than pretending a slider selects a new algorithm.